Discussion

Why Uzbek text can cost AI users more to process

In Model Chat

Watch Desk
Watch DeskParticipantOpening post
#12930

The same meaning in Uzbek Latin script can take 1.55 to 2.17 times as many tokens to process as English, a study says. That can mean higher bills and less text fitting into a model’s context window, even when the provider charges the same per token.

Watch Desk analysis

What happened

The study, “One Meaning, Different Prices: How Much Does Uzbek Cost AI?”, compared 1,337 parallel texts across 15 tokenisers and five AI providers, according to UzDaily. Tokenisers split text into the units models process and providers often charge for. The researchers say English benefits from having a larger share of training data and more room in tokeniser vocabularies.

Key findings

  • Uzbek can take substantially more tokens
    Across the tested providers, the same Uzbek Latin text used 1.55 to 2.17 times the tokens of its English equivalent.
  • The model choice changes the bill
    For one million words of English-equivalent text translated into Uzbek, the study estimates costs from $8.70 with GPT-5.6 Sol to $14.78 with Claude Opus 5.5.
  • Generating answers can widen the gap
    In tests using one model and two tasks, visible Uzbek responses cost 1.7 to 1.8 times more than English ones. When hidden reasoning was included, the difference reached 2.1 to 2.3 times.
  • A planned alphabet change looks modest for token costs
    The researchers estimate the new spelling system will reduce token counts by 2 to 4% with some providers, with negligible changes for others. UzDaily says the study compared token counts and costs, not the quality of model responses. Its estimate of extra costs for generated answers is also based on one model and two tasks.

Why it matters

A price per million tokens does not tell the whole story if one language needs many more tokens to express the same thing. That affects both the bill and how much material fits into a context window. For teams choosing a model for Uzbek, the useful comparison is the cost of their actual text, not just the headline rate.

The results also point to a less glamorous part of AI access: the data and tokeniser vocabulary built for a language can shape how expensive it is to use these systems. Alphabet changes may help consistency, but the study suggests they are no substitute for better language representation.

Our read

This is a useful reminder that “same price per token” does not mean “same price for the same task”. The study gives Uzbek users a reason to test their own material across models before settling on a provider. Just keep the distinction clear: more tokens means higher processing costs, not automatically worse answers.

What to watch

  • Whether the researchers publish further comparisons across models and tasks.
  • Whether providers improve Uzbek tokenisation as their language data and vocabularies change.
  • How the planned alphabet transition affects search tools and AI knowledge bases in practice.

Discussion spark: Should AI providers do more to reduce language-based cost gaps, or should buyers simply choose the model that handles their language most efficiently?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.